Preview AI models (also referred to as experimental models) are early-stage AI models that enable organizations to explore new capabilities before they are fully production-ready. These models are optional and intended for evaluation, testing, and experimentation.
What are preview AI models?
Preview AI models provide access to advanced capabilities ahead of general availability. These models help organizations evaluate emerging functionality, but they may have limitations in availability, reliability, performance, and support.
Preview and experimental models:
-
Are released before full general availability
-
May change frequently
-
May be removed or updated without notice
Important: Preview and experimental models are intended for exploration and testing and are not recommended for production use.
Model release types
AI models available in Microsoft Online Services can be categorized by release type:
-
Experimental
-
Designed for early experimentation
-
Not intended for production use
-
May have significant limitations in availability, reliability, or performance
-
-
Preview
-
Early access models that may evolve into generally available models
-
Not yet production-ready
-
-
Generally available (GA)
-
Fully released models intended for production scenarios
-
Provide established reliability, support, and performance
-
-
Default
-
The standard model selected for most scenarios
-
Typically the recommended GA model
-
Limitations of preview and experimental models
Preview and experimental models provide early access to new capabilities, but they also have important limitations:
-
Performance can vary, including response quality and latency
-
Models may be unavailable or experience interruptions
-
Features and behavior may change without notice
-
Availability may vary by region
Note: Use preview and experimental models only in test or evaluation scenarios.
Data handling considerations for models with data retention by the AI provider
Some preview and experimental models require data retention by the model provider.
When these models are enabled:
-
Prompts and responses may be processed and stored by the model provider
-
Data handling is governed by the model provider's terms and policies
-
Data may be processed outside your organization's geographic boundaries
Important: When a model requires data retention, the model provider, not Microsoft, controls how data is stored and processed.
Before enabling these models, review the provider's documentation and policies.
For example, Anthropic Claude Fable 5 is a Mythos-class model available as an optional Preview model with Data Retention in Microsoft 365 Copilot. Tenant admins can opt-in to enable access to this model from the Microsoft 365 Admin Center. Your organization's use of Claude Fable 5 and other Anthropic Preview models with Data Rentention is subject to Anthropic's Commercial Terms of Use, Data Protection Addendum, and data retention practices for Mythos-class models.
Admin controls for preview AI models
Administrators manage access to preview and experimental models using settings in the Microsoft 365 admin center.
Admins can:
-
Enable or disable preview models
-
Restrict access to specific users or groups
-
Configure access based on organizational requirements
Some models may require additional configuration, such as enabling access to external AI providers.
Models disabled by default
Some preview models are disabled by default, including:
-
Models with data retention by the provider (including Claude Fable 5)
Best practices
To safely evaluate preview AI models:
-
Enable preview models only for targeted users or test groups
-
Avoid using preview models in production workloads
-
Communicate data handling expectations to users
-
Review model provider documentation before enabling models with data retention
-
Monitor usage and adjust access as needed
